ipv4: fix race in concurrent ip_route_input_slow()

CPUs can ask for local route via ip_route_input_noref() concurrently.
if nh_rth_input is not cached yet, CPUs will proceed to allocate
equivalent DSTs on 'lo' and then will try to cache them in nh_rth_input
via rt_cache_route()
Most of the time they succeed, but on occasion the following two lines:
	orig = *p;
	prev = cmpxchg(p, orig, rt);
in rt_cache_route() do race and one of the cpus fails to complete cmpxchg.
But ip_route_input_slow() doesn't check the return code of rt_cache_route(),
so dst is leaking. dst_destroy() is never called and 'lo' device
refcnt doesn't go to zero, which can be seen in the logs as:
	unregister_netdevice: waiting for lo to become free. Usage count = 1
Adding mdelay() between above two lines makes it easily reproducible.
Fix it similar to nh_pcpu_rth_output case.

Fixes: d2d68ba ("ipv4: Cache input routes in fib_info nexthops.")
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Author: Alexei Starovoitov

net/ipv4/route.c

@@ -1776,8 +1776,12 @@ out:	return err;
 		rth->dst.error= -err;
 		rth->rt_flags 	&= ~RTCF_LOCAL;
 	}
-	if (do_cache)
-		rt_cache_route(&FIB_RES_NH(res), rth);
+	if (do_cache) {
+		if (unlikely(!rt_cache_route(&FIB_RES_NH(res), rth))) {
+			rth->dst.flags |= DST_NOCACHE;
+			rt_add_uncached_list(rth);
+		}
+	}
 	skb_dst_set(skb, &rth->dst);
 	err = 0;
 	goto out;